Learned foundation transfer104: source proposal
This is an unexecuted, fixed synthetic experiment. The source parent starts from
the pinned random initializer and learns by supervised gradients from raw x,y.
The previous portable quadrant result used authored hidden features; it is not
the foundation used here. No fitting, testing, importing or compilation occurred
while this proposal was authored.
The source classifies which coordinates exceed zero. The target classifies which
coordinates exceed x=0.2 and y=-0.2. Action indices retain their threshold-bit
meanings, while the thresholds differ deliberately. These are related tiny tasks;
the exact arithmetic teacher is simpler and perfect. This cannot establish useful
repair, Poker, general transfer, hostile isolation or scientific significance.
Fixed comparison
Three source initializer seeds (11,29,47) each produce an ordinary gradient-trained
parent and a sham parent trained on a predeclared permutation of the source labels.
The sham changes label association, preserves observations/class counts and uses
the same initialization, sampler and recipe. It is not a neuron permutation.
Five target arms use the same target rows, evaluation partition, replacement-draw
stream, 300 updates, batch16, learning rate0.01 and clipping1. Fullwarm starts with
all eight learned source tensors and trains all. Scratch trains all from the
random trunk and a fresh target head. Three head-only arms freeze the learned,
random or sham trunk respectively and share exactly the same fresh target policy
and value head. Consequently fullwarm has an intentional initial learned-head
advantage; its result does not isolate trunk transfer. The learned-vs-random and
learned-vs-sham head comparisons isolate the declared feature intervention more
closely. No arm or seed may be discarded, stopped early, tuned or replaced.
DATA.json contains every ordered row, including shuffled labels and their source
indices. All four coordinate grids are disjoint. Parent fitting reads sourceTrain
only; sham fitting reads shuffledSource only; target fitting reads targetTrain
only. Protected sourceAudit and targetEvaluation are inference observations, never
fit inputs or a selection signal. Target support is intentionally unequal under
shifted thresholds, so results must retain class counts beside aggregate accuracy.
Explicit semantic bridge
Ordinary checkpoint import refuses the source-to-target task binding. bridge.mjs
implements a new fixed experiment-only transformation, not an exception to that
guard. Source parents are minted into a private token table; expected content
identity is captured before ordinary source import. Copies write values into
fresh target-owned tensor arrays, preserve the destination binding and verify the
copied inverse. Source and target units, coordinate order/domain, action-index
meanings, source hashes and architecture are checked. Threshold configuration
changes intentionally. The runner supplies fixed tasks and data; there is no CLI
provider, model path, adapter callback or arbitrary layer-mask input.
The bridge's public JavaScript functions are local trusted composition helpers,
not security capabilities against arbitrary same-process callers. A digest minted
by this experiment is provenance rather than independent approval. Source receipt
fields check lineage consistency; they do not authenticate that a foreign caller
actually fitted a model. Root admission must bind the complete fixed runner and
all dependencies before any use. No generic checkpoint migration is authorized.
Retention has two distinct behavioral observations on protected source points:
the changed target model using original source labels, and a source-owned inference
recomposition containing the original source head with the target trunk. The latter
tests representation retention; it cannot establish retention of the changed target
policy. Every original parent remains unchanged. Recomposition retains the original
source model version tag, with its changed identity bound by the new content digest
and transformation receipt; it is not an optimizer-resume checkpoint.
Trainer and observations
trainer.mjs calls the actual reviewed numerical createAdam,
computeSupervisedGradients and applyAdam APIs with all/head-only. It makes one
fresh optimizer per complete fit; observation checkpoints never reset Adam. This
new wrapper does not fabricate a harness head-only trainingReceipt. Its own versioned
receipt binds encoded data, sampler state/draw count, scope, frozen weights/moments,
parent/result identity and options. Head clipping uses the selected head gradient
norm. Value tensors remain untrained. Full-sized gradient/moment allocations still
exist even in head-only mode.
The test source includes exact comparison of wrapper and existing fitTeam operation
order for both scopes, an independent uniform-softmax derivative/first-Adam scalar
calculation, partition truth, bridge refusals/inverses and crossing/censoring cases.
These tests are unexecuted and need a separate finite diagnostic selection. They
must not be counted as transfer evidence. Two equivalence declarations each perform
two 300-update fits; diagnostic costs are outside the experiment's8100-update vector.
Source observations occur at0,100,...600; target observations at0,20,...300. Training
never responds to them. Crossing means the first observed tau success, with a bracket
between observations, not the exact first successful update. No crossing is censored
above300. Receipt elapsed time includes scheduled inference and preservation checks,
but excludes initial observation, bootstrap and publication. Curve times are snapshot
start times and exclude the current snapshot's inference. They are descriptive and
cannot establish compute speedup. Whole-process CPU/wall/RSS require an external
owned-process controller and actual terminal evidence.
Qualification and resource boundary
PROTOCOL.json fixes utility, paired loss margins, retention margins and all seeds
before any fit. Thresholds are design choices, not normative ANN quotas. A passing
toy gate is an observation on these grids; all five original/general qualification
fields stay false. Source pretraining cost is reported separately and hypothetical
N=1,5,10 amortization is clearly identified as repetitions of this same target.
The source proposes one Node process under CPU30, active wall90, FD64, old heap128MiB,
16MiB per output file,64MiB physical output tree and observed RSS256MiB classifier.
It supplies no process controller and admits no execution. Root must independently
review source and diagnostic tests, select exact runtime/source pins and a new output,
and review a controller that retains interruption, raw streams and child reaping.
Node loader/startup/native/parent overhead and joint fit remain unknown. Cooperative
pre/post file checks are not a strong loaded-module identity or hostile-code boundary.
The normal success tree is64 payloads plus its manifest. Output writes include the
manifest bytes in the storage check. A late failure may leave a completed-looking
report beside a failure receipt or missing terminal evidence; root terminal and
independent readback govern completion. Filesystem interruption may prevent even a
failure receipt. Preserve all partials; never retry or change caps, seeds or gates in
place. Original historical ANN failures, SPENT scopes, full-scope/retention/resource
obligations and the actual E3 native training history are untouched.
The author previously contributed portable head-scope tests/source review. This is
source authorship, not independent validation. Existing numerical/harness files are
copied byte for byte with full source provenance; redistribution rights remain
unselected in the portable checkpoint. This folder is an owned local research
proposal, not a new publication/license grant.